Local Learning in Memristive Neural Networks for Pattern Reconstruction

Resistive devices such as memristors have attracted the researchers' attention as fundamental computing elements. The prime objective of this work is to provide a simulated analogue computing platform based on memristor devices and recurrent neural networks that exploits the memristor device conductance programmability to implement local learning algorithms. We present the application of two simple two-phase learning procedures for Dynamic Neural Networks used to solve a pattern reconstruction task. The first learning scheme is related to Energy-based models and the second generalizes this method to generic vector field dynamics, relaxing the requirement of an energy function. Experimental results show that both the two approach significantly outperforms conventional learning rules used for pattern reconstruction.